In a new data hall in western China, a technician walks past cabinets that are bolted down, cabled and switched off. The floor is polished. The cooling loops have been pressure-tested. Everything is ready except the demand.
The hall belongs to one of the country's largest cloud operators, and it is typical. Similar buildings are being commissioned across Guizhou, Inner Mongolia and Ningxia, sometimes before a single paying customer has signed for the capacity inside them.
That ordering — build first, fill later — is the defining financial fact of the buildout, and the hardest to see in a quarterly report.
Key takeaways
- Alibaba, Tencent and Baidu spent roughly US$20bn on capital projects in Q2 2026, more than double a year earlier (SemiAnalysis).
- All three ran negative free cash flow in that quarter, so the buildout is carried by balance sheets and cloud price increases.
- China's data-centre capacity should pass 24 GW by end-2026, against about 56 GW in the United States (SemiAnalysis).
- In March 2026, Tencent Cloud raised prices on some model inputs by 463%; Alibaba Cloud lifted AI compute and storage by up to 34%.
- National token volume hit 140 trillion a day by March 2026 (National Data Administration). Volume is not revenue.
The buildout is visible. The return is not
China's data-centre capacity is on course to exceed 24 GW by the end of 2026, against roughly 56 GW in the United States (SemiAnalysis, 25 September 2026). Compute-centre electricity demand is forecast near 300 TWh this year (Beijing Daily). The country operates 42 ten-thousand-card compute clusters (算力) (CAICT), and the planning ministry has pointed to more than RMB 7 trillion (≈ US$986bn / HK$7.7tn) of related investment in 2026 alone (NDRC, cited by CGTN).
All of that is capacity, not utilisation.
Who actually pays
Alibaba, Tencent and Baidu spent a combined US$20bn on capital projects in the second quarter of 2026, more than double the same period a year earlier, and all three reported negative free cash flow (SemiAnalysis). ByteDance, which is not listed, accounts for about one-fifth of delivered capacity, most of it leased rather than owned.
Read those together: the industry is buying compute faster than it sells it, and the gap is financed rather than earned. There are three sources, and only one is a customer.
The first is the cloud price list. In March 2026, Tencent Cloud raised prices on some model inputs by 463%, Alibaba Cloud lifted AI compute and storage by up to 34%, and Baidu Cloud moved by 5% to 30% (China Development Network; China Economic Net). The increases followed two years of falling token prices: DeepSeek V3.1 is listed at RMB 2 and RMB 8 per million input and output tokens (≈ US$0.3 / US$1.1), Doubao Lite at RMB 0.3 and RMB 0.6, and Qwen3.7-Max at US$2.5 and US$7.5 (CGTN, 18 June 2026).
The second is the balance sheet. Depreciation on a hall full of accelerators starts when the equipment is installed, not when it is sold. A three-year-old card depreciates at the same rate whether it serves one customer or none.
The third is municipal capital. Guangzhou alone runs a RMB 20 billion (≈ US$2.8bn / HK$22bn) industry fund, and city- and district-owned funds with an AI focus number nearly 200, holding more than RMB 160 billion (≈ US$22.5bn / HK$176bn) (Guangzhou municipal government, 14 April 2026). That money sits at the equity layer, absorbing risk that would otherwise reach a lender.
None of this is improper. Each moves the bill to a different year or balance sheet; none requires a Chinese enterprise to pay a price that covers its cost.
Three numbers that would settle the argument
Chinese operators disclose capacity, clusters and tokens, not what a card earns. If that changes, read these:
- Utilisation. A 24 GW national figure shows what could be switched on; the share of installed accelerators running paying inference shows what the buildout is worth.
- Cloud gross margin by segment. Blended margins hide the practice of selling AI infrastructure below cost to win workloads that then stay.
- Resale prices for older cards. A rising secondary market suggests scarcity; a falling one suggests the installed base already exceeds demand.
Agents are the reason to expect demand to keep rising anyway. Agentic coding consumes 10 to 50 times the tokens of plain chat (CGTN, 18 June 2026), and national token volume rose from about 100 billion a day in early 2024 to 140 trillion by March 2026 (National Data Administration). That is a real demand curve. The open question is whether it is profitable at RMB 0.3 per million tokens.
The constraint that decides the ending
The physical ceiling is memory, not concrete. CXMT's 2026 output of about 2.2 million HBM stacks supports only 250,000 to 400,000 Ascend-class packages, while SMIC can produce die for more than a million (SemiAnalysis estimate). That mismatch explains why Huawei's card prices rose sharply in 2026, with its rotating chairman saying supply is short and priority goes to domestic demand.
Scarcity is what allows price increases to stick. If memory supply loosens, the March rises become harder to defend and the depreciation bill arrives with less revenue behind it. If it tightens, customers pay more for less compute, and the buildout looks less like overcapacity than rationing.
Either way, the deciding variable is not the model leaderboard. It is how much of the installed base is working, and at what price.
Honest limitations
The capex and cash-flow figures are SemiAnalysis estimates rather than audited disclosures, and company capital spending includes non-AI projects. No Chinese operator publishes accelerator-level utilisation or AI infrastructure revenue, so the central claim rests on capex, cash flow and price increases together. The March price changes were reported by Chinese industry press and apply to selected products, not every SKU. The cluster count measures announced clusters, not utilisation.
What readers can do now
- Operators: when a Chinese cloud quotes an unusually low price per million tokens, ask which accelerator generation serves it and for what committed term; the discount usually prices in hardware already depreciating.
- Investors: read cloud segment margin and free cash flow together, not capex alone; capex growth without margin recovery is a subsidy, whichever balance sheet holds it.
- Researchers: treat token volume as a demand proxy and nothing more, and ask for utilisation data before concluding anything about Chinese compute capacity.